Journal of Neuroscience Methods 274 (2016) 94105

Contents lists available at ScienceDirect

Journal of Neuroscience Methods

journal homepage: www.elsevier.com/locate/jneumeth

The inverse problem in electroencephalography using the bidomain

model of electrical activity
Alejandro Lopez Rincon , Shingo Shimoda
RIKEN BSI-TOYOTA Collaboration Center, 2271-130 Anagahora, Shimoshidami, Moriyama-ku, Nagoya, Aichi 463-0003, Japan

h i g h l i g h t s
The EEG inverse problem is solved using the bidomain model.
A spatial comparison is made with fMRI using the linkRBrain platform.
Accuracy is increased in comparison with other methods (MNE and LORETA).

a r t i c l e

i n f o

Article history:
Received 17 May 2016
Received in revised form
28 September 2016
Accepted 28 September 2016
Keywords:
EEG
Bidomain
Regularization
Inverse problem

a b s t r a c t
Background: Acquiring information about the distribution of electrical sources in the brain from electroencephalography (EEG) data remains a signicant challenge. An accurate solution would provide an
understanding of the inner mechanisms of the electrical activity in the brain and information about
damaged tissue.
New Method: In this paper, we present a methodology for reconstructing brain electrical activity from EEG
data by using the bidomain formulation. The bidomain model considers continuous active neural tissue
coupled with a nonlinear cell model. Using this technique, we aim to nd the brain sources that give rise
to the scalp potential recorded by EEG measurements taking into account a non-static reconstruction.
Comparison with Existing Methods: We simulate electrical sources in the brain volume and compare
the reconstruction to the minimum norm estimates (MNEs) and low resolution electrical tomography
(LORETA) results. Then, with the EEG dataset from the EEG Motor Movement/Imagery Database of the
Physiobank, we identify the reaction to visual stimuli by calculating the time between stimulus presentation and the spike in electrical activity. Finally, we compare the activation in the brain with the registered
activation using the LinkRbrain platform.
Results/Conclusion: Our methodology shows an improved reconstruction of the electrical activity and
source localization in comparison with MNE and LORETA. For the Motor Movement/Imagery Database,
the reconstruction is consistent with the expected position and time delay generated by the stimuli. Thus,
this methodology is a suitable option for continuously reconstructing brain potentials.
2016 The Author(s). Published by Elsevier B.V. This is an open access article under the CC BY license
(http://creativecommons.org/licenses/by/4.0/).

1. Introduction
Neural processes are generated by the propagation of electrical activity in the brain. This activity produces electrical potentials
that can be measured through electrodes in various positions on the
scalp, a technique referred to as electroencephalography (EEG). This
voltage distribution on the scalp is generated from the extracellular current by the post-synaptic potentials in the apical dendrites
of pyramidal neurons inside the brain. EEG signals, in comparison

Corresponding author.
E-mail addresses: alejandro.lopezrn@hotmail.com (A. Lopez Rincon),
shimoda@brain.riken.jp (S. Shimoda).

with other brain imaging techniques, have the advantage of high

temporal resolution, but they have a small amplitude (on the order
of hundred of V) and are highly susceptible to noise.
The electrical activity of the brain is described by the volume
conductor model with current sources using Poissons equation
coupled with Neumann and Dirichlet boundary conditions (Hallez
et al., 2007a). Simulating the potentials at the electrode positions
from current sources inside the brain is known as the EEG forward
problem; inference of the position of the current sources from electrode potentials is known as the EEG inverse problem or the neural
source imaging problem (Grech et al., 2008; Brannon et al., 2008).
The EEG inverse problem is fundamental in neuroscience, as it
gives insight about spatial and temporal activity in the brain for
different tasks. An accurate solution of the neural source imaging

http://dx.doi.org/10.1016/j.jneumeth.2016.09.011
0165-0270/ 2016 The Author(s). Published by Elsevier B.V. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).

A. Lopez Rincon, S. Shimoda / Journal of Neuroscience Methods 274 (2016) 94105

problem can contribute to understanding the inner workings of the

brain and to pinpointing regions with conductivity anomalies that
might indicate damaged tissue (Pascual-Marqui, 1999). The EEG
inverse problem is an ill-posed problem; thus, there is not a unique
solution. To reconstruct an approximate solution, we need regularization techniques and methods like minimum norm estimates
(MNE) (Grech et al., 2008) and low resolution electrical activity
tomography (LORETA) (Grech et al., 2008; Pascual-Marqui et al.,
2002, 1999). These methods consider the relationship between the
current sources and the measured potentials assuming a quasistatic approximation expressed by the lead eld matrix (Weinstein
et al., 1999).
In this work, we propose to solve the EEG inverse problem
by using the bidomain model (Sundnes, 2007). The bidomain is a
reaction-diffusion model for the electrical activity of the heart and
takes into account the anisotropy of the intracellular and extracellular cell domains. Compared with other methods, it does not
impose a quasi-static assumption and considers an electrical model
of a cell described by a series of ordinary differential equations. The
bidomain model is typically used to describe the hearts electrical
activity, but it was adapted as an alternative method to solve the
EEG forward problem in Yin et al. (2013) and Szmurlo et al. (2007).
Starting from the standard bidomain formulation, we coupled
the model to the node lead eld matrix and created the necessary
operators to solve the inverse problem, which gives a relationship
between the scalp potentials and the stimuli in the cell model.
Compared with other source localization methods, the bidomain
method maintains the continuum assumption. Instead of applying
regularization techniques to the current sources, we apply the regularization to the stimuli that produce the current sources. This is
similar to the approach explained in detail in Lopez-Rincon et al.
(2015), but adapted to the brain.

Ii =

To explain the bidomain formulation, it is necessary to give a

brief overview of the lead eld matrix, MNE, and LORETA methods
for the EEG source localization problem.
2.1.1. The lead eld matrix
The EEG-measured neural activity from the brain can be
described by Poissons equation for electrical conduction (De
Munck et al., 1988; Weinstein et al., 1999)
(1)

(2)

where  represents the electrostatic potentials,  the conductivity,

I the current sources in the brain volume, n the outward normal
vector,  the surface area of the head, and  the volume of the head.
In EEG modeling, we consider the normal component of the current
density to be zero as a boundary condition. Using nite element
method (FEM) discretization in a 3D mesh, we can write Eqs. (1)
and (2) as a system of linear equations, which may be written in
matrix form (Sundnes, 2007; Gockenbach, 2006):
(3)

where


i j ,

Aij =


and u is a vector with the scalar node values of the potential, for
basis functions i and j (Fig. 1).
The model described in Eq. (1) is known as the pure Neumann
problem and has no unique solutions; however, applying additional constraintsfor example, reducing it to Laplaces equation
(Johnson and MacLeod, 1998), xing one electrode on the scalp to
zero (Becker et al., 1982; Troparevsky and Rubio, 2003) or using the
method described in Bochev and Lehoucq (2005)gives a unique
solution. From the system in Eq. (3) we can construct the lead eld
matrix L which gives a projection between the current sources in
the brain volume and the measured electrical activity in the scalp:
r = L s + noise.

with the boundary condition

Au = I

Ii ,


2.1. Mathematical background of bidomain formulation

  n = 0 on ,

Fig. 1. Top: 3D mesh of a head divided into tetrahedra using FEM. In this geometry, the head is the domain  and the outer surface is the domain . Bottom: FEM
discretization of the domains.

2. Methods

  = I in ,

95

(4)

Here, L is the lead eld matrix, r is a vector of the measured potentials on the head, and s a vector of the current sources in the brain
volume. For our tests, we use the node lead eld matrix as described
in Weinstein et al. (1999) and thereby reconstruct not only the
current sources, but also the potential in the brain volume. The
lead eld matrix will typically be non-invertible as it depends on
the quantity of sources and recordings. Thus, it is necessary to use
regularization methods to solve the inverse problem
minLs r.
s

(5)

2.1.2. Minimum norm estimates

The MNE (Grech et al., 2008) method is suitable for reconstructing the activity on the cortical surface. This method will give the
minimum energy solution (closest to zero). MNE does not have an
inclusion of priori restrictions that allow approximating a solution
closer to the actual physical behavior in the brain from the set of

96

A. Lopez Rincon, S. Shimoda / Journal of Neuroscience Methods 274 (2016) 94105

possible solutions (Vogel, 2002; Wang et al., 2011). Eq. (5) in terms
of the brain sources is equal to
s = [LT L + I]

1 T

L r.

(6)

where I is the identity matrix, and  the regularization parameter.

To choose the regularization parameter we use the L-curve algorithm (Hansen and OLeary, 1993). The L-curve is a parametric plot
of (log10 (LT Ls LT r  2 ), log10 (s  2 )) for different values of the
regularization parameter . The optimal value of  for Tikhonov0 regularization can be obtained from the maximum value of the
curvature given by
() =


 

(( )2 + (
)2 )

(7)

3/2

where
= log10 (LT Ls LT r2 )

(8)

= log10 (s 2 )

(9)

2.1.3. Low resolution electrical tomography

LORETA is similar to the MNE method but, instead of using the
identity matrix for the regularization, it uses the discrete Laplace
operator. LORETA is suited to smoothly distributed sources as it
takes into account the connectivity given by the discrete Laplacian
(D ) (Grech et al., 2008; Pascual-Marqui et al., 2002, 1999). In this
case the solution is
s = [LT L + TD D ]

1 T

L r.

2.1.4. Bidomain formulation for brain activity propagation

The bidomain model assumes that electrical activity is generated by the depolarization of the cell membrane between the
intracellular and extracellular domains (Sundnes, 2007; Henriquez,
1992). The bidomain approach was developed to describe to electrical activity in cardiac tissue, but since then it has been adapted
to other systems, such as the brain (Szmurlo et al., 2006, 2007; Yin
et al., 2013). This model has the advantage of combining the overall electrical activity of the brain and the discrete nonlinear cell
model. For each point in the mesh it solves the cell model and then,
through diffusion, calculates the overall electrical activity of the
brain. The bidomain model is described by the following equations
(Tung, 1978):

(Mi v) + (Mi ue ) = (Cm

v
+ Iion (v, w) + Iapp ),
t

(Mi v) = ((Mi + Me ) ue ),

(11)
(12)

where v is the transmembrane potential, ue is the extracellular

potential, Mi is the intracellular conductivity tensor, Me is the extracellular conductivity tensor, Cm is transmembrane capacitance, and
is the membrane surface to volume ratio. If we assume a linear relationship between the intra- and extracellular conductivity
tensor, Me = Mi , where is a constant scalar, we can reduce the
bidomain model to a monodomain model:

v
(Mi v) = (Cm
+ Iion (v, w) + Iapp ),
1+
t

(13)

with the boundary condition

(Mi v) n = 0,

(Mi v) = ((1 + )Mi ue ).

(15)

We will scale the equations with

Mi
.
Cm

to simplify the notation. The bidomain and monodomain models

are dependent on the ber direction for approximating the propagation of electrical activity. To create the ber directions, we use
the white matter in the brain mesh and draw the ber directions
tangent to the surface taking into account the anisotropy in white
matter (the anisotropy in gray matter is negligible) (Hallez et al.,
2007b) giving the ber directions represented in Fig. 2.
2.1.5. The monodomain inverse operator
From Eq. (13) we create the operators to solve the inverse
problem using a method similar to that explained in detail in
(Lopez-Rincon et al., 2015). To solve the monodomain model
numerically, we use the Godunov operator splitting technique
(Sundnes, 2007) to divide the system into an ionic part (Eq. (17))
and a diffusion part (Eq. (18)):

v
= Iion (v, w) Iapp ,
t

(17)

v

(Mi v).
=
1+
t

(18)

Then, we rst solve

sp1
= Iion (v, w) Iapp ,
t

(19)

sp1 (0) = v(0)

(20)

for t [0, t]. This gives us sp1 (t). Next, we solve

(14)

and the following equation to get the extracellular potential

Mi =

Fig. 2. Lateral and superior views of the brain mesh showing the ber directions.

(10)

sp2

(Mi (x) sp2 ),
=
1+
t

(21)

sp2 (0) = sp1 (t)

(22)

for t [0, t] to get sp2 (t) which we set equal to

(16)

v(t) = sp2 (t).

(23)

A. Lopez Rincon, S. Shimoda / Journal of Neuroscience Methods 274 (2016) 94105

97

Fig. 3. Superior and lateral views of the EEG electrode positions.

Thus, we can solve the system in steps. If we discretize Eq. (17) over
time, we obtain

vn+1 vn
t

= Iion (vn , wn ) Iapp ,

(24)

or

vn+1 = tIion (vn , wn ) tIapp + vn .

(25)
Fig. 4. BEM Laplace interpolation of electrode values.

Eq. (18) is discretized using the  rule:

vn+2 vn+1
=
t

1+

(Mi vn+2 )

(1 )

(Mi vn+1 ) .
1+

Multiplying by a test funcion

vn+2

 t
1+

= vn+1

and rearranging the terms, we have

+

H

vn+1

 t  
1+
(1 )

Mi vn+1 n.


Mi vn+2

t
1+

vn+1
j,
j

(29)

N


vn+2
j
j

(30)

j=1

(27)

where N is the number of nodes in the brain volume. Thus, by

applying the FEM approximation, Eq. (28) becomes

N


Mi vn+2 n

N

j=1

vn+2 =

t
(Mi vn+1 )
1+

H

vn+1 =

Applying Greens identity to Eq. (27) gives

vn+2

(26)

(Mi vn+2 )

+ (1 )

We will consider vn+1 and vn+2 as linear combinations of basis functions:


vn+2
(
j

j=1

Mi vn+1


=
(28)

N

j=1

j i + (


t
)
1+


vn+1
(
j

j i (1 )


Mi j i )


t
1+


Mi j i ),

i, j = 1, ..., N

(31)

98

A. Lopez Rincon, S. Shimoda / Journal of Neuroscience Methods 274 (2016) 94105

and we get the following matrices.

Aij =

j i + 

 t  
1+

t
Bij =
j i (1 )
1
+


Mi j i ,

(32)


Mi j i .

(33)

>From Eqs. (32) and (33), we can construct the matrix equation
AVn+2 = BVn+1 ,

(34)

where Vn+2 and Vn+1 are the vectors with the nodal values for iterations n + 1 and n + 2 in the numerical solution. From Eq. (15) we
derive a relationship between the transmembrane potential and
the extracellular potential, given by
RVn+2 = QUe .

(35)

From Eqs. (34) and (35) we can explicitly construct a vector with
the nodal values for the extracellular potential Ue as
Ue = Q1 RA1 BVn+1 .

A relationship with the voltage distribution over the scalp can be

constructed using the lead eld matrix L multiplied by a Dirichletto-Neumann operator (A1
) to transform currents in the brain
bb
volume into potentials T = LA1
.
bb

TQ1 RA1 BVn+1 = r,

(37)

or
TQ1 RA1 B(tIion tIapp + Vn ) = r,

(38)

where Iapp and Iion are vectors with the nodal values for the applied
current and ionic ux. From this we will create the operator
P = TQ1 RA1 B,

PtIion PtIapp + PVn = r.

(40)

The EEG inverse problem is an ill-posed problem, making a regularization technique necessary. We used the following Tikhonov
functional:
min(|| PtIapp r PtIion ||2 + ||C(Iapp Iapp  )||2 ),  > 0,(41)
Iapp

or
min(|| Pts r PtIion ||2 + ||C(s s )||2 ),
s

 > 0,

(42)

with the L-Curve method (Hansen and OLeary, 1993) to nd the

regularization parameter. Here C is a constrained matrix (the identity matrix), and s is the prior information (s = 0). For our tests, we
use  = 0.05.
2.1.6. Error estimation
The error is given by the difference between the original potential distribution and the solution obtained by each method, as in
Lopez-Rincon et al. (2015):
N Nodes


i
i,j

2.2. Implementation
For the implementation, we created several modules using the
described methods in the C# programming language. For the example of real data, we implemented an EDF reader for the Physionet
Physiobank database (Schalk et al., 2004; Goldberger et al., 2000).
We created a module in OpenGL to visualize and modify 3D meshes
to create the ber directions for the method and visualize the electrical activity. The mesh used in all the examples consists of 35,982
elements. For the visualization of the results we added a plug-in to
output the results in Gmsh format (Geuzaine and Remacle, 2009)
(Fig. 3).

(39)

and

error =

Fig. 5. Simulated spike values on the head.

(36)

i,j

i,j

|Ue(original) Ue(calculated) |

(43)

where Ue( ) are the nodal values of the potential for the original
and reconstructed distributions, Nodes is the maximum number of
nodes, and N is the total number of samples.

2.2.1. Data preprocessing

To use the EEG signals we need to preprocess the data. The
rst step is to allocate the electrode signals to a node in the
head mesh. The EEG signals were recorded from 64 electrodes
following the international 10-10 system (Schalk et al., 2004).
Then, for each sample in the dataset, with N being the total
number of samples we solve the Laplace equation using the boundary element method (BEM) (Schlitt et al., 1995; Nintcheu Fata,
2009) with the following conditions: each given electrode position will be considered a Dirichlet boundary condition, and the
rest will be considered a null-ux Neumann condition. The result
is shown in Fig. 4. Then, we make a second interpolation in time to
match the time step to be used in the bidomain formulation (t).
Finally, we normalize the data. This procedure is summarized in
Algorithm 1.
Algorithm 1.

Preprocessing of the data

A. Lopez Rincon, S. Shimoda / Journal of Neuroscience Methods 274 (2016) 94105

99

Fig. 6. Measured spike values on the head from a real EEG measurement.

2.2.2. The bidomain inverse algorithm

Once we have the interpolated values on the scalp and the operators described in the last section, we can solve the inverse problem
using the bidomain formulation according to Algorithm 2.
Algorithm 2.

Bidomain inverse

Fig. 7. Original distributions for the forward problem with one source at 50 ms and
200 ms.

3.1. Comparison of the MNE, LORETA, and bidomain methods for

one source

3. Results
3.0.1. Example of the forward problem
Using the procedure for the bidomain model described in
Szmurlo et al. (2006), we created electrical activity in the scalp.
The EEG simulated data are shown in Fig. 5, and the measured
scalp values from the EEG Motor Movement/Imagery Dataset of
the Physiobank are shown in Fig. 6 (Schalk et al., 2004; Goldberger
et al., 2000).

We put one source in the brain volume and construct the potentials in the scalp . Then, we solved the inverse problem with MNE
(Fig. 8), LORETA (Fig. 9) and the bidomain approach (Fig. 10). In each
case, we compare the solution to the original distributions (Fig. 7)
at 50 ms and 200 ms.
3.2. Comparison of the MNE, LORETA, and bidomain methods for
three sources
We put three sources in the brain volume and constructed the
potentials in the scalp . Then, we solved the inverse problem with
MNE (Fig. 12), LORETA (Fig. 13) and the bidomain approach (Fig. 14).
In each case, we compared the solution to the original distribution
Fig. 11.

100

A. Lopez Rincon, S. Shimoda / Journal of Neuroscience Methods 274 (2016) 94105

Fig. 9. LORETA solution for one source at 50 ms and 200 ms.

Fig. 8. MNE solution for one source at 50 ms and 200 ms.

Table 2
Comparison of reconstruction times in ms for each test.

Table 1
Comparison of errors for each test.

One source
Three sources

MNE

LORETA

Bidomain

11.2e4
16.9e4

41.4e3
67.4e3

31.1e3
65.4e3

MNE
LORETA
Bidomain operator

One point

Three points

7110906
10308247
13510721

7111203
10311428
13511285

To make a comparison, we calculate the error given by Eq. (43)

for each method. These errors are summarized in Table 1.

3.3. Response to visual stimuli

3.2.1. Time comparison

We next compare the different methods using a computer with
an Intel(R) Core(TM) i7-2860QM CPU @ 2.50Hz and 8.00 Gb RAM.
The times in ms for reconstruction programmed using C# and
OpenGL for the graphics are shown in Table 2.
This time is dependent of the processor in the computer; therefore, the values here should be taken as giving only an overall idea.

The EEG Motor Movement/Imagery Dataset (Schalk et al., 2004)

from the Physiobank (Goldberger et al., 2000) has a series of measurements for different tasks. We used the recordings where a
visual stimulus appears on a screen and then the subject closes
his/her st. The measurements originated from recordings made
with a 64-channel EEG system (BCI2000).
Using the bidomain formulation, we solved the EEG inverse
problem to reconstruct the response to visual stimuli in the brain

A. Lopez Rincon, S. Shimoda / Journal of Neuroscience Methods 274 (2016) 94105

Fig. 10. Bidomain solution for one source at 50 ms and 200 ms.

volume. From the different subjects, we obtained the timings

between the stimulus being received in the brain and the reaction to it. The exact time depends on the subject, but on average
remains the same. In Fig. 15, the addition of the absolute value of
the electrical activity in the brain (vertical axis) is shown for two
subjects in two different tasks for half a second (horizontal axis).
Measuring the time between the centers of the rst and second highest peaks for 45 reconstructions gives an average of
0.246666667 s, which is consistent with the timing of the visual
stimulus being detected in the observers brain before he/she manually responds to the stimulus (Amano et al., 2006).
Comparing the position of the second spike shown Fig. 17 with
data from the platform LinkRbrain, which is a collection of data that
shows the greatest activity in the fMRI bold signal during several
tasks, we nd that the results are consistent with hand movement
(Fig. 16).

101

Fig. 11. Original distributions for the forward problem with three sources at 50 ms
and 200 ms.

4. Discussion
First we simulated a spike in the brain using the forward bidomain method, as given in (Szmurlo et al., 2006; Yin et al., 2013), to
show that the system works.
Then we made a comparison of the MNE, LORETA, and bidomain approaches to the inverse problem using synthetic data
for one and three cortical sources. We chose to compare our
approach with MNE because it is a method suited for reconstructing sources on the cortical surface with and LORETA because it
gives satisfactory results in EEG source analysis (Grech et al., 2008).
From a comparison of the original distribution with the ones created with the cost functional in Eq. (41) and with the ones from
MNE and the LORETA model, by visual inspection of the relevant
images, it is clear that using the dynamic model gives more precise
results.

102

A. Lopez Rincon, S. Shimoda / Journal of Neuroscience Methods 274 (2016) 94105

Fig. 12. MNE solution for three sources at 50 ms and 200 ms.

Fig. 13. LORETA solution for three sources at 50 ms and 200 ms.

As is well known, EEG source localization is an ill-posed problem. Our approach attacked this problem by modeling a EEG
dynamic changes computing from a dynamics of neuron activities.
The observations of EEG dynamic changes can change EEG source
localization problem to a well-posed problem. This is a remarkable
point of our approach that distinguish our method from the conventional approaches with the quasi-static models and possible to
provide more precise results as mentioned above.
We used real measurement data and calculated the inverse solution for three subjects undertaking 15 tasks each, to estimate the
time delay between the rst two highest activity peaks. We reconstructed the electrical activity in the brain for the Physiobank EEG
Motor Movement/Imagery Dataset and used measurements from
64 electrodes on the scalp for the st-closing action (Schalk et al.,
2004; Goldberger et al., 2000). When these data were collected, the
subjects received a visual stimulus and then closed their sts. In the
reconstructed activity, the rst spike is considered as brain activity

generated from the visual stimulus and the second as the response
to it.
To validate our results, we compared the spatial representation of the highest activity with LinkRbrain (Mesmoudi et al.,
2015). LinkRbrain is an open-access web platform for multi-scale
data integration and visualization. The functional part of this tool
(300 sensorimotor/cognitive functions) was reconstructed from the
fMRI literature ( 5000 papers). For our experiments, we choose the
hand movement task in the database as a reference and compared
the highest activity peaks with the reconstructed inverse solution
from the bidomain formulation. The positions of the peaks are consistent with those in the database. Finally, we compared the time
delay between the reception of the visual stimulus and the intention of movement for the hand movement task. The time delays are
consistent with what is found in literature for the visual stimulus
response and hand movement task.

A. Lopez Rincon, S. Shimoda / Journal of Neuroscience Methods 274 (2016) 94105

Fig. 16. Activity registered for hand movement in the LinkRBrain database.

Fig. 14. Bidomain solution for three sources at 50 ms and 200 ms.

Fig. 15. Overall activity in the brain from 2 subjects for hand movement.

103

104

A. Lopez Rincon, S. Shimoda / Journal of Neuroscience Methods 274 (2016) 94105

However, it should be remembered that we are working with

archived measurements and not readings from the original brain
and head mesh of subjects. With the necessary pre-processing and
a powerful enough computer, this method could be performed in
real time, as can be done with MNE or LORETA. In future work,
we will divide the EEG measurements into different frequency
bands to see if we are able to classify movement intention from
the inverse solution.
Acknowledgments
This work was conducted as part of a collaboration between the
RIKEN Brain Research Institute and Toyota Motor Company.
References

Fig. 17. Activity reconstructed by solving the bidomain inverse problem.

5. Conclusion
In this paper, we applied the bidomain formulation to the
inverse problem of brain activity source localization from EEG
signals. The inverse bidomain formulation allows us to take into
account the non-linearity of the cell models, time propagation, and
the anisotropy in the conductivity of the brain volume tissue, in
comparison to other methods.
The method makes a construction of the sources considering the
dynamic natural depolarization and not the quasi-static approach
as other methods (Baillet et al., 2001; Gener and Williamson, 1998;
Pascual-Marqui, 1999). In other words, the proposed method uses
the solution at time n to construct the solution at time n + 1, not
only the EEG signals at time n + 1. The results show that this method
gives a better reconstruction for source localization in comparison
with the quasi-static approach.
In an example of visual stimulus and response, the reconstruction appears to be temporally and spatially close to reality.

Amano, Kaoru, et al., 2006. Estimation of the timing of human visual perception
from magnetoencephalography. J. Neurosci. 26 (15), 39813991.
Baillet, S., et al., 2001. Evaluation of inverse methods and head models for EEG
source localization using a human skull phantom. Phys. Med. Biol. 46 (1), 77.
Becker, E.B., et al., 1982. Finite elements, an introduction. J. Appl. Mech. 49, 682.
Bochev, Pavel, Lehoucq, Richard B., 2005. On the nite element solution of the pure
Neumann problem. SIAM Rev. 47 (1), 5066.
Brannon, Elizabeth M., et al., 2008. Principles of Cognitive Neuroscience. Vol. 83.
Sinauer Associates, Sunderland, MA.
De Munck, Jan C., Van Dijk, Bob W., Spekreijse, Henk, 1988. Mathematical dipoles
are adequate to describe realistic generators of human brain activity. Biomed.
Eng., IEEE Transactions on 5 (11), 960966.
Gener, Nevzat G., Williamson, Samuel J., 1998. Differential characterization of
neural sources with the bimodal truncated SVD pseudo-inverse for EEG and
MEG measurements. IEEE Trans. Biomed. Eng. 45 (7), 827838.
Geuzaine, Christophe, Remacle, Jean-Fran?ois, 2009. Gmsh: A 3-D nite element
mesh generator with built-in pre-and post-processing facilities. Int. J. Numer.
Methods Eng. 79 (11), 13091331.
Gockenbach, Mark S., 2006. Understanding and Implementing the Finite Element
Method. Siam.
Goldberger, A.L., Amaral, L.A.N., Glass, L., Hausdorff, J.M., Ivanov PCh, Mark, R.G.,
Mietus, J.E., Moody, G.B., Peng, C.-K., Stanley, H.E., 2000 June 13. PhysioBank,
PhysioToolkit, and PhysioNet: Components of a New Research Resource for
Complex Physiologic Signals. Circulation 101 (23), e215e220 [Circulation
Electronic Pages; http://circ.ahajournals.org/cgi/content/full/101/23/e215].
Grech, Roberta, et al., 2008. Review on solving the inverse problem in EEG source
analysis. J. Neuroeng. Rehabil. 5 (1), 1.
Hallez, Hans, et al., 2007a. Review on solving the forward problem in EEG source
analysis. J. Neuroeng. Rehabil. 4 (1), 1.
Hallez, Hans, et al., 2007b. Importance of including anisotropic conductivities of
grey matter in EEG source localization. In: 6th International Conference on
Bioelectromagnetism (ICBEM2007). Vol. 9. No. 2.
Hansen, Per Christian, OLeary, Dianne Prost, 1993. The use of the L-curve in the
regularization of discrete ill-posed problems. SIAM J. Sci. Comput. 14 (6),
14871503.
Henriquez, Craig S., 1992. Simulating the electrical behavior of cardiac tissue using
the bidomain model. Crit. Rev. Biomed. Eng. 21 (1), 177.
Johnson, Christopher R., MacLeod, Robert S., 1998. Adaptive local regularization
methods for the inverse ECG problem. Progr. Biophys. Mol. Biol. 69 (2),
405423.
Lopez-Rincon, Alejandro, Bendahmane, Mostafa, Ainseba, BedrEddine, 2015. On 3D
numerical inverse problems for the bidomain model in electrocardiology.
Comput. Math. Appl. 69 (4), 255274.
Mesmoudi, Salma, et al., 2015. LinkRbrain: Multi-scale data integrator of the brain.
J. Neurosci. Methods 241, 4452.
Nintcheu Fata, Sylvain, 2009. Explicit expressions for 3D boundary integrals in
potential theory. Int. J. Numer. Methods Eng. 78 (1), 3247.
Pascual-Marqui, Roberto D., et al., 2002. Functional imaging with low-resolution
brain electromagnetic tomography (LORETA): a review. Methods Findings Exp.
Clin. Pharmacol. 24 (Suppl C), 9195.
Pascual-Marqui, Roberto D., et al., 1999. Low resolution brain electromagnetic
tomography (LORETA) functional imaging in acute, neuroleptic-naive,
rst-episode, productive schizophrenia. Psychiatry Res.: Neuroimaging 90 (3),
169179.
Pascual-Marqui, Roberto D., 1999. Review of methods for solving the EEG inverse
problem. Int. J. Bioelectromagnetism 1 (1), 7586.
Schalk, G., McFarland, D.J., Hinterberger, T., Birbaumer, N., Wolpaw, J.R., 2004.
BCI2000: A General-Purpose Brain-Computer Interface (BCI) System. IEEE
Trans. Biomed. Eng. 51 (6), 10341043 [In 2008, this paper received the Best
Paper Award from IEEE TBME.].
Schlitt, Heidi A., et al., 1995. Evaluation of boundary element methods for the EEG
forward problem: effect of linear interpolation. Biomed. Eng. IEEE Transactions
on 2 (1), 5258.
Sundnes, Joakim, et al., 2007. Computing the Electrical Activity in the Heart. Vol. 1.
Springer Science & Business Media.

A. Lopez Rincon, S. Shimoda / Journal of Neuroscience Methods 274 (2016) 94105

Szmurlo, R., et al.,2006. Bidomain formulation for modeling brain activity
propagation. In: Electromagnetic Field Computation, 2006 12th Biennial IEEE
Conference on. IEEE.
Szmurlo, R., et al.,2007. Multiscale nite element model of the electrically active
neural tissue. EUROCON, 2007. In: The International Conference on &# 34;
Computer as a Tool &# 34;. IEEE.
Troparevsky, M.I., Rubio, D., 2003. On the weak solutions of the forward problem in
EEG. J. Appl. Math. 2003 (12), 647656.
Tung, Leslie, 1978. A bi-domain model for describing ischemic myocardial dc
potentials. Diss. Massachusetts Institute of Technology.

105

Vogel, Curtis R., 2002. Computational Methods for Inverse Problems. Vol. 23.
Siam.
Wang, Yanfei, Yagola, Anatoly G., Yang, Changchun, 2011. Optimization and
Regularization for Computational Inverse Problems and Applications. Higher
Education Press, Beijing.
Weinstein, David, Zhukov, Leonid, Johnson, Chris, 1999. Lead eld basis for fem
source localization. In: Frontiers in Simulation, Simulationstechnique 18th
Symposium.
Yin, S., Dokos, S., Lovell, N.H., 2013. Bidomain Modeling of Neural Tissue. Neural
Engineering. Springer US, 389404.